Legacy Post: Artificial intelligence and the environment: Putting the numbers into perspective

 

This is a legacy post, aimed at our members and discussing information that was widely available at the time. We leave posts that have been superseded in place as a historical record.

For the latest information on AI and the environment, we recommend the United Nations University’s Environmental Cost of AI Energy Use (June 2026)

Michael Webb

 

Headlines around generative AI’s environmental impact often focus on isolated numbers and statistics, which can appear alarming without context. However, some recent articles, such as ‘What’s the impact of artificial intelligence on energy demand?’ written by Hannah Ritchie and Andy Masley’s ‘Using ChatGPT is not bad for the environment’, have started to look at these numbers in context and place these impacts alongside those of other digital activities. While we haven’t verified the exact numbers in these articles, they do seem plausible. This blog summarises some of the key findings and highlights other interesting research we’ve come across.  We believe that making these kinds of comparisons can help to clarify the broader environmental impact of generative AI.

In some of our recent communities, our members have mentioned the importance of digital sustainability education. Not everyone is aware of how much energy, water and carbon their online activities are using, but the reality is ‘everything we do digitally involves the vast transfer of data through the internet from one place to another, brokered through datacentres’.

Energy

Although exact numbers for ChatGPT’s energy consumption are not yet available, a reasonable estimate is around 2.9 Wh per request. This assumption is based on data reporting that a typical Google search reportedly uses around 0.3 Wh, and interacting with an LLM could ‘likely cost ten times more than a standard keyword search’.

From public data available online, we found the following:

• Asking an LLM (like ChatGPT) a single query uses around 0.0029 kWh
• Streaming Netflix in HD for one hour uses around 0.077 kWh

The figure of 0.077 kWh for streaming Netflix in HD dates back to 2020, but it’s unlikely to have changed significantly since then, and we believe it’s still a reasonable estimate. From these numbers alone, watching one hour of Netflix is equivalent to approximately 26.5 ChatGPT queries. For comparison, I’d estimate I ask around five ChatGPT queries a day and stream between one – two hours of Netflix (or other streaming platforms) each day. Based on energy use alone, my streaming habits likely have a much higher impact. That being said, it isn’t always this simple as the numbers are sensitive to the device you use, the type of network connection you have, and which resolution you choose. This comparison is also dependent on how often someone uses ChatGPT or spends time streaming. It ultimately comes down to individual habits, which can vary widely.

In his research, Andy Masley also calculated that streaming Netflix and YouTube videos consume far more energy than ChatGPT, yet these figures aren’t as prevalent as generative AI’s energy use in the media. You can read how he calculated these figures specifically in his Substack post, but he makes a great point that just because these services use a lot of energy, doesn’t mean we should stop using them. Instead, ‘what matters when considering what to reduce, is the energy used compared to the amount of value produced’. We think this is a great way of approaching all digital activities, especially those of generative AI.

Water

Data centres consume water for various online activities, with some having a higher impact than others. While we don’t currently have any official figures for specific AI models, a research paper estimated that GPT3 uses 500ml of water for every 10 to 50 medium length responses, depending on when and where it is deployed. It’s important to highlight that this study didn’t just measure the water used for each response, but the number also included the water consumed for the training of the model and the data centre’s operational water footprint as well, suggesting that trying to pinpoint an exact amount is difficult. The findings in Masley’s article suggested that a single ChatGPT query actually has a relatively low water footprint (30ml) compared to streaming music (250ml), spending one hour browsing social media (430ml) and joining a one hour Zoom call (1720ml). While a single query may not use much water, the cumulative impacts of course add up. However, this can also be said about Zoom calls and browsing social media.

Another study found that one hour of streaming or videoconferencing requires 2-12 litres of water. If we use Masley’s calculations that asking ChatGPT a question uses around 30ml of water, then asking the average five queries a day, still only equates to about 150ml which is far less than that required for videoconferencing. While these numbers are only estimations, it highlights the point that we don’t necessarily question being on a video call and using our camera for an hour, but some of us do question the use of LLMs for tasks that add real value to our work and personal life.

Carbon emissions

The carbon footprint of generative AI is an active area of discussion, and once again, depends on a multitude of factors. HuggingFace encourages developers to calculate the carbon emissions of their models, and advise that key factors include the type of computing hardware, where and how often the model is used, training time and the local energy mix, particularly the use of renewables versus non-renewables. As we don’t have the exact information from big tech, estimating the carbon footprints of popular LLMs can be challenging.

Masley did some calculations for the USA and factored in ‘both the energy used in data centres and the energy used on each individual device’. He estimated that Spotify, Netflix and YouTube use far more energy than ChatGPT per American household, but also points out that streaming has replaced more energy-intensive physical processes from the past (e.g. physically producing CDs). He concluded that while reducing your use of ChatGPT is an option, asking ChatGPT 50,000 fewer questions likely has only a minor impact on overall emissions compared to other activities. Ultimately, while our personal choices matter, their effectiveness is largely influenced by broader policy trends.

In the UK, 238g of CO₂ is emitted per kWh of electricity generated. If we take the figure from above (2.9Wh) and convert it to kWh, we can estimate that one LLM query uses about 0.69 grams of CO₂. If we increased that to five queries, we’re looking at around 3.45 grams of CO₂. We can compare this to a 2020 analysis by the International Energy Agency, which found that streaming for one hour emits roughly the same amount of CO₂ as boiling a kettle once (34g). Again, there are a lot of variables to consider when it comes to figures, but without the transparent data from tech companies, we can only refer to these estimates. As Masley mentions, ‘even if you are only focused on lifestyle changes, it is best to focus on the most impactful lifestyle changes for climate’ and turning away from all your digital activities isn’t the quick fix, nor is it particularly a viable option given the world we live in.

Estimations

Based on publicly available data, we’ve compiled the estimates into the table below. While these figures are approximate and not definitive, we think they offer a reasonable snapshot of the environmental impact. It may be useful to calculate your own average daily use and compare with other digital activities, such as streaming.

It’s important to note that all of this does not undermine the impact generative AI is having, and we certainly don’t think any of this is intended to make people think using LLMs makes no difference to the environment, because it does. There are many ways to reduce our carbon footprint, but simply avoiding generative AI won’t solve the problem on its own.  Given that companies aren’t sharing their specific numbers, we would be better off critically evaluating the numbers we read about and question how they fit into the wider context. In our next blog, we’ll look ahead at developments in this area and some initiatives taking place.

---

Are you currently promoting sustainable and responsible use of AI in your institutions? Have you come across any other interesting comparisons? We’d love to hear from you.

If you work in the FE and Skills sector, you may be interested in joining Jisc’s FE and Skills Digital Sustainability Community. If so, please fill out the expression of interest form via this link and we will email you the invite to join the Teams channel. This online community space will be dedicated to fostering a collaborative environment where members can share ideas, resources, and best practices related to digital sustainability.

Find out more by visiting our Artificial Intelligence page to view publications and resources, join us for events and discover what AI has to offer through our range of interactive online demos.

Join our AI in Education communities to stay up to date and engage with other members.

Get in touch with the team directly at AI@jisc.ac.uk